Generative AI tools like ChatGPT are entering classrooms quickly, but research on their long-term effects on learning is mixed. Cognitive psychology distinguishes fast, automatic System 1 thinking from effortful System 2 processing—the latter is essential for deep learning. Using AI to bypass difficult cognitive work can yield short-term gains yet undermine long-term understanding and produce metacognitive errors. Well-designed AI tutors that scaffold effort and prompt active engagement may help, but they cannot replace the hard cognitive work required for mastery.
How AI Is Changing How We Learn — Why Hard Work Still Matters
Learning and mastery come from effort, whether that's done with a powerful chatbot or AI tutor or not, but educators and students need to resist outsourcing that work.Francesco Carta fotografo via Getty Images
When OpenAI introduced a "study mode" for ChatGPT in July 2025, company representatives claimed the feature could improve students' academic performance. That promise has educators and researchers asking: is this marketing speak, or does evidence support the claim?
What the Early Research Shows
Generative AI is entering classrooms rapidly, but rigorous long-term research on its effects lags behind adoption. Some early studies report benefits for particular groups—such as computer programming students and English-language learners—and a May 2025 paper in Nature suggested chatbots might support learning and higher-order thinking. Still, many experts have flagged methodological weaknesses in parts of this literature.
Other studies raise concerns. Some research finds that greater reliance on ChatGPT during study correlates with poorer performance on later tasks when the AI is not available. In short, early evidence is mixed: AI can sometimes help, but it can also enable shortcuts that weaken durable learning.
System 1 and System 2: A Useful Framework
Cognitive psychologists often distinguish two modes of thought. System 1 is fast, intuitive and automatic—useful for routine actions such as getting dressed or riding a bike. System 2 is slow, effortful and deliberate—the cognitive mode responsible for learning, reasoning, and forming durable knowledge.
Mastery depends on System 2. Struggle, friction and mental effort are not nuisances; they encode information, strengthen neural connections and enable transfer of knowledge to new tasks. You can't build reliable System 1 expertise without sustained System 2 practice first.
Think of the brain like a muscle: it grows stronger through effort. If you let a machine do the demanding work for you, the muscle atrophies.
When AI Becomes A Shortcut
Using AI to bypass difficult cognitive work—having it write essays, solve quizzes or supply ready-made arguments—removes the mental workout students need to build long-term understanding. Habitual cognitive offloading can impair learning and memory and produce metacognitive errors (misjudging what you actually know).
For example, a study found students who researched a topic via ChatGPT experienced lower cognitive load during the task but produced weaker reasoning about the subject than students who used traditional web searches. In another experiment, students who used AI to revise essays often scored higher immediately—frequently by copying AI-generated sentences—but showed no greater knowledge gains or transfer and engaged less in deep System 2 processes. Authors described this pattern as "metacognitive laziness": short-term gains at the expense of long-term skill development.
When AI Can Help
AI is not inherently harmful to learning. Offloading can be productive once foundational knowledge is in place. Well-designed AI tutors—models that ask probing questions, offer hints, scaffold practice and push learners to think—can act as scalable, individualized coaches.
Companies are pursuing this vision. In addition to OpenAI's study mode, Anthropic introduced a "learning mode" for Claude in April 2025 intended to engage learners in Socratic dialogue rather than simply providing answers. Early trials of tutor-like modes show promise but also new complications. One high-school math study found students who used the basic ChatGPT performed worse on a later exam without AI access; those who used a hint-focused tutor variant did no better than students who studied without AI, yet many mistakenly believed they had improved.
Practical Guidance For Educators And Students
Design and use matter. To amplify learning rather than enable shortcuts, AI tools should be configured to:
- Encourage active retrieval and problem solving rather than only giving answers,
- Scaffold practice and prompt reflection to reduce metacognitive errors,
- Be integrated into assessment designs that require unaided demonstration of skill, and
- Be paired with explicit instruction about when and how to use AI responsibly.
Conclusion
Like earlier transformative technologies—the smartphone, the internet or even writing—understanding AI's full effects on cognition will take time. The likely outcome is nuanced: AI's impact will depend on context, task design and how learners engage with it. What remains clear from decades of learning research is that deep knowledge and mastery require effortful System 2 practice. AI can amplify and support that work, but it cannot replace the mental effort that produces enduring learning.
Author: Brian W. Stone, Boise State University.
Disclosure: Brian W. Stone reports no conflicts of interest, consultancy, shareholdings or funding tied to companies that would benefit from this article.
This article is republished from The Conversation, a nonprofit, independent news organization.
Help us improve.
